Google DeepMind released AlphaGenome 2 on November 11, a model that predicts how a stretch of DNA, and any variant in it, behaves inside a specific human cell type: a heart muscle cell, say, or an immune cell responding to infection. Its predecessor could only predict broad tissue- or cell-line-level averages. DeepMind says the new model finds nearly twice as many variants that change gene expression as the previous state of the art.
Why it matters: most disease-linked genetic variants sit outside genes, in regulatory DNA whose effects often show up only in particular cells. Genome-wide association studies flag regions, not culprits. A model that can say which variant matters, in which cell, and in which direction is a shortcut to deciding which lab experiments are worth running. And unlike many lab announcements, this one is usable today: non-commercial researchers can query it through an API for free.
What changed from AlphaGenome 1
According to DeepMind's post, the first AlphaGenome read a one-million-letter DNA sequence and produced a fixed set of parallel outputs for bulk tissues and cell lines, which spread its capacity thin and made closely related cell types hard to tell apart. AlphaGenome 2 is a conditional model: you give it the DNA plus one condition (a tissue, a cell type or a cell state), and it predicts for that condition only.
- Readouts: gene expression, histone modifications and chromatin accessibility, as before, plus DNA methylation, which is new.
- Training data: human and mouse bulk and single-cell data; DeepMind cites more than 100 million single-cell profiles across 68 tissues (the body of its post says over 107 million).
- Bring your own cells: a cell encoder lets scientists define a condition from their own single-cell experiment. DeepMind says that from a few dozen cells of scRNA-seq data, it matched or beat models trained or fine-tuned directly on those datasets, without retraining.
- Filling gaps: given one measurement, such as single-cell gene activity, it can predict others that weren't measured, such as chromatin accessibility.
The numbers DeepMind reports
All figures below are DeepMind's own and come from its announcement; they haven't been independently replicated.
| Benchmark (TenK10K, 28 immune cell types) | AlphaGenome 2 | AlphaGenome 1 fine-tuned on single-cell data |
|---|---|---|
| Causal expression variants identified | 30% | 16% |
| Correct direction of gene effect | 43% | 23% |
DeepMind also points to two case studies. In inflammatory bowel disease, the model reproduced a published finding that a distant regulatory variant overactivates the ETS2 gene in gut macrophages, singling out the causal variant and showing the effect was confined to those cells. In data from a COVID-19 mRNA vaccine study, it separated vaccine-activated T cells from resting ones and indicated that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells.
Who can use it, and how
| Use | Access |
|---|---|
| Non-commercial research | Available now via the AlphaGenome API, with Colab tutorials |
| Commercial research | Private preview |
| Agent workflows | "AlphaGenome 2 Skills" for Google Antigravity, coming soon |
| Precomputed results | Predictions for trait-associated variants across hundreds of human cell and tissue conditions, covering traits such as height and type 2 diabetes |
DeepMind is also taking proposed use cases until November 30, with selected projects eligible for extra quota and credits. The post doesn't mention downloadable model weights or commercial pricing. On the industry side, Genomics plc CEO Peter Donnelly said the company plans to build the model into its Mystra platform, according to DeepMind.
Sarah Teichmann of the University of Cambridge, a co-founder of the Human Cell Atlas, highlighted that it "adapts directly to new single-cell data without retraining," as quoted by ShiftDelete and SiliconANGLE.
The limits
- Most causal variants are still missed. A 30% hit rate on the immune benchmark is close to double the old model's, but it means the majority go unidentified, as ShiftDelete notes.
- Correlation, not mechanism. ShiftDelete notes the model learns correlations rather than proven cause and effect, and SiliconANGLE stresses its predictions are meant to guide experiments, not replace them.
- Cost. DeepMind researcher Žiga Avsec told The Verge the model is more expensive to run, so it won't soon replace the original AlphaGenome or its genome-wide Atlas of all 9 billion possible single-letter changes, according to ShiftDelete.
- Research only. DeepMind says neither AlphaGenome version is validated or approved for clinical or diagnostic use.
Who should care: human geneticists sitting on GWAS hits they can't explain, single-cell labs that want variant-level predictions for their own cell states, and drug-discovery teams choosing targets. The practical next step is the free API tier and the November 30 use-case call; anyone hoping to use it on patients will have to wait for independent validation that doesn't exist yet.
Sources
- Google DeepMind — AlphaGenome 2: Deciphering how cells read DNA to better understand disease deepmind.google
- SiliconANGLE — Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model siliconangle.com
- ShiftDelete.Net — AlphaGenome 2: DeepMind's DNA model now works cell by cell shiftdelete.net